Abstract:
This thesis presents a comprehensive exploration of level-sets applied to various stages of image analysis, aiming to enhance understanding, modelling, and interpretability of image data. The research focuses on three critical aspects namely, data cleaning, data modelling, and explainability. In data cleaning, the adaptive median filter is a commonly used technique removing noise from images which compares individual pixels to an adaptive window around it. Herein the adaptive median filter is improved by acting on level-sets rather than individual pixels. The proposed level-sets adaptive median filter demonstrates effective noise removal while preserving edges in the images better than the traditional adaptive median filter. Secondly, this work considers representing images as graphical models, with the nodes corresponding to the fuzzy level-sets of the images. This novel representation successfully preserves and maps critical image information required for understanding of image context in a binary classification scenario. Further, this representation is used to propose a novel method for modelling images, which enables inference to be applied on image content directly. Finally, within the realm of deep learning object detection saliency maps, the detector randomised input sampling for explanation (D-RISE) is extended using informative level set sampling. A key, yet computationally expensive, component of the former is the generation of a suitable number of masks. The proposed methodology in this work, namely the adaptive D-RISE, harnesses proportional level-sets sampling of masks to reduce the required number of masks and improves the convergence of attribution.
